Park integrated energy system energy management method and system considering photovoltaic prediction error

Abstract

The invention relates to a park integrated energy system energy management method and system considering photovoltaic prediction errors, a photovoltaic digital twinborn prediction model based on data driving is constructed, and the core of the model is to construct a virtual twinborn body which is in real-time mapping with a physical photovoltaic system and is dynamically updated in parallel. Firstly, historical output data and meteorological data of a photovoltaic system are collected, a long-short-term memory neural network is used for learning time sequence characteristics of the data, an initial data driving prediction model is constructed, and tedious physical mechanism modeling is avoided. And then, monitoring output data of the physical entity in real time, dynamically comparing the output data with a model predicted value, and when an error exceeds a set threshold value, automatically triggering a model updating process, and carrying out online adjustment and retraining on hyper-parameters of the LSTM model. The mechanism ensures that the prediction model can be continuously self-optimized along with environment change and equipment state evolution, so that the robustness and reliability of the prediction precision in the whole life cycle are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of energy optimization and dispatching technology, and in particular to an energy management method and system for a comprehensive energy system in a park that takes into account photovoltaic prediction errors. Background Technology

[0002] Park Integrated Energy System (PIES) is a regional energy solution that integrates multiple energy sources such as electricity, heat, cooling, and gas, enabling efficient cascade utilization of energy and on-site consumption of clean energy.

[0003] However, the integrated energy system in the park faces numerous challenges in actual operation. In particular, photovoltaic power generation is significantly affected by external environmental factors such as weather, climate, and irradiance, exhibiting marked volatility and randomness. This uncertainty directly impacts the safe and stable operation and economic dispatch efficiency of the park's integrated energy system. Therefore, improving the accuracy of photovoltaic output forecasting and, based on this, achieving optimized dispatch that balances economy and robustness is one of the core issues currently of concern to the academic and engineering communities.

[0004] In photovoltaic (PV) forecasting, existing technologies mainly include physical forecasting methods and statistical forecasting methods. Physical forecasting methods require a large number of parameters and involve a complex modeling process. Statistical forecasting methods rely on historical data and machine learning models to predict PV output, but once the models are trained, they are difficult to adapt to long-term environmental changes, and the prediction accuracy may degrade over time.

[0005] In terms of optimal scheduling, existing research has proposed methods such as fuzzy optimization, stochastic optimization, robust optimization, and distributed robust optimization. In fuzzy optimization, the construction of fuzzy membership functions relies on subjective experience; when the optimization result is sensitive to the membership function, different construction methods may lead to different results. Stochastic optimization heavily relies on the probability distribution of uncertain parameters; if the distribution is inaccurate, the optimization result deviates from reality. Robust optimization always considers the worst-case operating conditions during modeling, often resulting in overly conservative optimization results. Scheduling strategies tend to sacrifice economy for safety, leading to relatively high operating costs. In the scenario of integrated energy island operation in a park, simply using robust optimization methods may result in underutilization of energy equipment, reduced absorption rates of renewable energy sources such as photovoltaics, and poor economic performance. Furthermore, robust optimization cannot reflect the probability differences of different uncertainties; that is, regardless of the probability of extreme situations occurring, they are considered with equal weight, leading to deviations between the optimization results and actual operational needs.

[0006] In summary, there is currently a lack of an energy management method and system for integrated energy systems in industrial parks to solve or partially solve the aforementioned problems. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology and provide an energy management method and system for a comprehensive energy system in a park that takes into account photovoltaic prediction errors, so as to solve or partially solve the problems of photovoltaic prediction models lacking real-time dynamic correction capabilities, difficulty in maintaining accuracy over a long period of time, and cumbersome modeling process and insufficient stability.

[0008] The objective of this invention can be achieved through the following technical solutions: One aspect of the present invention provides an energy management method for a comprehensive energy system in a park that takes into account photovoltaic prediction errors, comprising the following steps: Acquire the actual power output data of the photovoltaic array and the corresponding meteorological data to construct a training dataset; A data-driven prediction model based on a neural network is constructed. The data-driven prediction model takes meteorological data as input and output data as output, and is trained based on the training dataset. Acquire real-time meteorological data and actual power output data, use the trained data to drive the prediction model to obtain predicted power output data, and calculate the power output error between the actual power output data and the predicted power output data. Determine whether the processing error meets the preset conditions. If not, adaptively update the data-driven prediction model. If so, overlay the output error distribution on the predicted output data to generate an original scene set that covers possible changes in photovoltaic output in various future periods and conforms to the error distribution. Then, reduce the scene by clustering. Based on the reduced scenario set, a PIES distributed robust scheduling model is constructed under the constraints of the park's integrated energy system to solve start-up, shutdown, and power output decisions, thereby realizing energy management of the park's integrated energy system.

[0009] As a preferred technical solution, the PIES distributed robust scheduling model takes minimizing the system operating cost and the cost of wasting light as its objective function.

[0010] As a preferred technical solution, the constraints of the integrated energy system of the park include electrical power balance constraints, thermal power balance constraints, conventional unit constraints, and energy storage constraints. The conventional unit constraints include electrical power constraints, ramp-up constraints, and start-up and shutdown constraints. The energy storage constraints include upper and lower limit constraints and start-up and end-state constraints.

[0011] As a preferred technical solution, the PIES distributed robust scheduling model is as follows: in, , for , The feasible region consists of the corresponding constraints. These are deterministic quantities, unaffected by photovoltaic prediction errors, and include the start-up and shutdown status of conventional units and the charging and discharging status of electrical and thermal energy storage. This refers to the amount that is adjusted in real time according to changes in photovoltaic output, including the output value of each unit. For the cost of starting and stopping, For the cost of power generation and the cost of curtailment, The probability of the scenario occurring. Let be the set of feasible regions for the probability distribution of the scene. The total number of scenes, This is a transpose.

[0012] As a preferred technical solution, the solution process of the PIES distributed robust scheduling model includes the following steps: Under the constraints of 1-norm and ∞-norm, the start-stop and output decisions are obtained by solving the main problem based on the probability distribution of the initial output error. Under the premise of fixed start-stop and output decisions, the adversarial subproblem is solved in the ambiguity set to obtain the worst probability distribution and corresponding cost. If the difference between the objective values ​​of the main problem and the adversarial subproblem exceeds the tolerance, the worst distribution information is transformed into new constraints or columns and fed back to the main problem. The iteration continues until convergence.

[0013] As a preferred technical solution, the clustering is k-means clustering.

[0014] As a preferred technical solution, the 1-norm and ∞-norm constraints are modeled as follows: In the formula, and Representing 1-norm and - Permissible deviation of probability under the norm condition. , The confidence level represents the probability of uncertainty. The total number of scenes, To estimate the sample size for the photovoltaic prediction error distribution, The main problem is modeled as follows: In the formula, , for , The feasible region consists of the corresponding constraints. These are deterministic quantities, unaffected by photovoltaic prediction errors, and include the start-up and shutdown status of conventional units and the charging and discharging status of electrical and thermal energy storage. This refers to the amount that is adjusted in real time according to changes in photovoltaic output, including the output value of each unit. For the cost of starting and stopping, For the cost of power generation and the cost of curtailment, The probability of the scenario occurring. The total number of scenes, For transpose, To estimate the lower bound of the optimal value of the original problem, This represents the current iteration number. The adversarial subproblem is modeled as follows: In the formula, Let be the set of feasible regions for the probability distribution of the scene. The optimal values ​​of the decision variables in the first stage, obtained from solving the main problem, are... For a given first-stage decision Below, the feasible region of the decision variables in the second stage.

[0015] As a preferred technical solution, the adaptive update includes incremental training, transfer learning, or parameter correction.

[0016] As a preferred technical solution, the data-driven prediction model is constructed based on long short-term memory neural networks, recurrent neural networks, or convolutional neural networks.

[0017] As a preferred technical solution, the meteorological data includes solar irradiance, ambient temperature, and humidity.

[0018] Another aspect of the present invention provides an energy management system for a park integrated energy system that considers photovoltaic prediction errors, for implementing the aforementioned energy management method for a park integrated energy system, the system comprising: The data acquisition module is used to acquire the actual output data of the photovoltaic array and the corresponding meteorological data to build a training dataset; The data-driven model building module is used to build a data-driven prediction model based on a neural network. The data-driven prediction model takes meteorological data as input and output data as output, and is trained based on the training dataset. The interaction and correction module is used to acquire real-time meteorological data and actual output data, use the trained data to drive the prediction model to obtain predicted output data, calculate the output error between the actual output data and the predicted output data, and determine whether the processing error meets the preset conditions. If not, the data-driven prediction model is adaptively updated. If so, the output error distribution is superimposed on the predicted output data to generate an original scene set that covers possible changes in photovoltaic output in various future periods and conforms to the error distribution. Scene reduction is performed through clustering. The decision-making module is used to construct a PIES distributed robust scheduling model under the constraints of the park's integrated energy system based on the reduced scenario set, solve start-up and power output decisions, and realize energy management of the park's integrated energy system.

[0019] Compared with the prior art, the present invention has at least one of the following beneficial effects: (1) Improving the accuracy of wind power output prediction: The data-driven prediction model constructed in this invention effectively captures the nonlinear mapping relationship between the temporal characteristics of photovoltaic power output and meteorological factors by introducing a neural network as the core prediction unit, thus avoiding the dependence on precise parameters in traditional physical modeling. Simultaneously, by constructing a real-time data interaction and dynamic correction mechanism, the digital twin model can continuously compare the predicted values ​​with the actual output. When the error exceeds a threshold, it automatically triggers the model update process, enabling online adjustment or incremental learning of model hyperparameters. This invention not only improves the initial accuracy of prediction but also effectively combats model degradation through a closed-loop feedback mechanism, providing a real-world photovoltaic power output data foundation for the subsequent optimized scheduling of the park's integrated energy system.

[0020] (2) The scheduling strategy avoids overly conservative approaches while ensuring stability: This invention searches only for the worst-case probability distribution within the probability ambiguity set, rather than a single worst-case scenario, thereby effectively reducing system operating costs while ensuring robustness. Meanwhile, traditional robust scheduling ignores probabilistic information, while distributed robust scheduling introduces... / The norm-constrained probability deviation set allows for flexible adjustment of the probability distribution within permissible limits, enhancing adaptability to uncertainty and ensuring the robustness of scheduling results. Finally, distributed robust scheduling, based on photovoltaic digital twin prediction results, generates scenarios and constructs a probability ambiguity set, which better reflects the true statistical characteristics of photovoltaic power output, making it more engineering feasible and practically valuable compared to traditional robust scheduling. Attached Figure Description

[0021] Figure 1 This is a flowchart of the energy management method for the integrated energy system of the park in the embodiment; Figure 2 This is a schematic diagram of the digital twin framework of photovoltaics in the PIES example; Figure 3 This is a schematic diagram of the overall logic of distributed robust scheduling of PIES in the embodiment; Figure 4 This is a schematic diagram of the energy management system of the integrated energy system in the park, as shown in the embodiment. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0023] Example 1 To address the problems of the aforementioned existing technologies, this embodiment provides an energy management method for a park's integrated energy system that considers photovoltaic (PV) prediction errors. This method constructs a data-driven PV digital twin prediction model. The core of this model lies in building a virtual twin that is mapped in real-time to the physical PV system and dynamically updated in parallel. First, historical output data and meteorological data of the PV system are collected. A Long-Short Term Memory (LSTM) neural network is used to learn its temporal characteristics, constructing an initial data-driven prediction model, avoiding cumbersome physical mechanism modeling. Then, by monitoring the physical entity's output data in real-time and dynamically comparing it with the model's predicted values, when the error exceeds a set threshold, the model update process is automatically triggered, adjusting and retraining the LSTM model's hyperparameters online. This mechanism ensures that the prediction model can continuously self-optimize with environmental changes and equipment state evolution, thereby significantly improving the robustness and reliability of prediction accuracy throughout its entire lifecycle.

[0024] Specifically, the methods include Figure 2 The digital twin framework (corresponding to steps S1-S4) and Figure 3 The distributed robust scheduling (corresponding to steps S4-S5) consists of two parts.

[0025] See Figure 2 This invention presents a photovoltaic (PV) digital twin framework for an integrated energy system in a park. Based on a Long Short-Term Memory (LSTM) neural network, the framework achieves rolling PV power prediction by fusing historical output data with real-time meteorological data. It also incorporates error modeling and uncertainty quantification mechanisms, outputting a power sequence containing the predicted mean and probability distribution. This embodiment further utilizes residual distribution to generate and reduce typical scenarios, providing input for distributed robust optimization scheduling. During operation, the digital twin framework features online monitoring and adaptive updates. When prediction accuracy decreases or the external environment changes, it can trigger incremental model training or retraining, thereby ensuring long-term stable prediction performance. Through the above technical solutions, this invention achieves dynamic mapping between PV predictions and the actual system state, providing support for the safety, economy, and robustness of the integrated energy system in an isolated operation environment.

[0026] See Figure 1 The photovoltaic digital twin framework includes the following steps: Step S1: Obtain the actual power output data of the photovoltaic array and the corresponding meteorological data to construct a training dataset.

[0027] The physical entity layer collects real-time data on the actual power output of the photovoltaic array, as well as meteorological data affecting its power generation efficiency (such as solar irradiance, ambient temperature, humidity, etc.). The historical database stores a large amount of historical meteorological data and corresponding historical power output data, providing a data foundation for model training.

[0028] Step S2: Construct a data-driven prediction model based on a neural network. The data-driven prediction model takes meteorological data as input and output data as output, and is trained based on the training dataset.

[0029] A data-driven prediction model is established based on a long short-term memory neural network (LSTM). This model can effectively learn the temporal characteristics, intraday periodicity and seasonality of photovoltaic power output, as well as the complex nonlinear mapping relationship between meteorological factors and power output.

[0030] Furthermore, other neural networks can also be used to build data-driven prediction models: 1) Recurrent Neural Network (RNN). The gated recurrent unit (GRU) is similar to LSTM, both of which are based on gating mechanisms to process time series data [9]. In photovoltaic power output prediction, GRU can effectively capture intraday periodicity and short-term meteorological changes, and can also achieve real-time prediction and dynamic correction.

[0031] 2) Convolutional and Deep Feature Extraction. Convolutional Neural Networks (CNNs) extract local features from weather data and historical power output curves through convolutional layers, effectively capturing spatial correlations and local temporal patterns. The resulting CNN-RNN hybrid model can be used for short-term photovoltaic forecasting and scene generation.

[0032] Step S3: Obtain real-time meteorological data and actual power output data, use the trained data to drive the prediction model to obtain predicted power output data, and calculate the power output error between the actual power output data and the predicted power output data.

[0033] During the online operation phase of the digital twin, the framework inputs real-time data into a trained LSTM model to obtain the predicted output force. Simultaneously, the system continuously acquires the actual output force of the physical entity and compares the two in real time to calculate the prediction error.

[0034] Step S4: Determine whether the processing error meets the preset conditions. If not, adaptively update the data-driven prediction model. If yes, proceed to the subsequent distributed robust scheduling.

[0035] The system determines whether the error meets the preset accuracy requirements. If it does, the current twin model is accurate and reliable, and the prediction results can be directly output for subsequent optimization and scheduling. If it does not meet the requirements, the model's adaptive update mechanism is triggered. This mechanism retrains or incrementally learns the model by adjusting its hyperparameters or injecting new real-time data, enabling the digital twin model to evolve dynamically and maintain predictive performance highly consistent with the physical entity.

[0036] See Figure 3 This is a schematic diagram of distributed robust scheduling, which specifically includes the following steps: Step S4: After processing whether the error meets the preset conditions, based on the predicted power output data and its power output error distribution, multiple samples are extracted in the multidimensional random space using the Latin hypercube sampling method to generate an original scene set covering possible changes in photovoltaic power output in various future periods, and scene reduction is performed through clustering.

[0037] This step utilizes historical and real-time meteorological data (including irradiance, temperature, humidity, etc.) collected within the park, along with actual photovoltaic (PV) output data, to construct a data-driven prediction model based on a Long Short-Term Memory (LSTM) network. This model obtains predicted PV power values ​​for each future scheduling period. Simultaneously, the digital twin model operates synchronously with the actual physical system, monitoring the deviation between the predicted values ​​and actual output in real time. When the deviation exceeds a preset threshold, an adaptive update mechanism is triggered, dynamically correcting the model through incremental training, transfer learning, or parameter adjustment to ensure the long-term validity and reliability of the prediction results. In this way, the digital twin model not only reflects the nonlinear characteristics of the PV system under different operating environments but also rapidly adjusts to changes in the external environment, enabling real-time evolution of the prediction model. This provides accurate data support for subsequent generation of uncertain scenarios and robust scheduling optimization.

[0038] This step, based on the photovoltaic (PV) predictions and their error distribution obtained from the digital twin model, employs the Latin hypercube sampling (LHS) method to extract a large number of samples in a multidimensional random space, generating an original scene set covering possible changes in PV output over future periods. This ensures uniform coverage of the probability distribution space with a limited number of samples. Then, to reduce computational complexity, this embodiment uses the k-means clustering algorithm for scene reduction.

[0039] By superimposing an error distribution on the photovoltaic power forecast, multiple scenarios that conform to the error distribution are generated using scenarios, and the scenarios are reduced to a few typical scenarios and their corresponding probabilities.

[0040] Step S5: Based on the reduced scenario set, construct a PIES distributed robust scheduling model under the constraints of the park's integrated energy system, solve the start-stop and output decisions, and realize the energy management of the park's integrated energy system.

[0041] The reduced set of scenarios provides initial values ​​for the solution process, with the objective function being the minimization of system operating costs and curtailment costs. Constraints include electrical power balance constraints and thermal power balance constraints. Conventional generating units need to meet corresponding electrical power constraints, ramp-up constraints, and start-up / shutdown constraints. Energy storage needs to meet charge / discharge power constraints, upper and lower limits on energy storage capacity, and initial and final state constraints. Photovoltaic power output and electric boiler output need to meet upper and lower limit requirements.

[0042] Therefore, the basic PIES scheduling model can be expressed as: In the formula: It is a deterministic quantity, unaffected by photovoltaic prediction errors, including the start-up and shutdown status of conventional units and the charging and discharging status of electrical and thermal energy storage. This refers to the amount that is adjusted in real time according to changes in photovoltaic power output, including the output value of each unit. For start-stop costs, For the cost of power generation and the cost of curtailment, This is the photovoltaic prediction vector. The coefficient matrix, To be related to the decision variables in the first stage The relevant constraint coefficient matrix, Let be the constant vector on the right-hand side of the inequality constraint. Let be the constant vector on the right-hand side of the equality constraint. This is the vector of predicted photovoltaic output values.

[0043] Based on this, a distributed robust model is constructed, which combines robust optimization methods and stochastic optimization methods. It obtains the scheduling scheme by seeking the worst-case probability distribution of photovoltaic power output. The two-stage distributed robust model can be modified from the basic scheduling model as follows: The first stage in the formula (the stage of solving the main problem) To avoid changing according to actual scenarios. This represents the quantity of change in the second stage (the stage of solving adversarial subproblems). It primarily involves the probability distribution in the worst-case scenario to find an optimal value that meets the objective. In the formula, , for , The feasible region consists of the corresponding constraints. These are deterministic quantities, unaffected by photovoltaic prediction errors, and include the start-up and shutdown status of conventional units and the charging and discharging status of electrical and thermal energy storage. This refers to the amount that is adjusted in real time according to changes in photovoltaic output, including the output value of each unit. For the cost of starting and stopping, For the cost of power generation and the cost of curtailment, The probability of the scenario occurring. Let be the set of feasible regions for the probability distribution of the scene. The total number of scenes, This is a transpose.

[0044] Based on the scene set reduced in step S4, the reduced scene set provides initial values ​​for the solution process in order to find the worst-case probability distribution of the scene. Based on the initial value, add the 1-norm and - Norm has two constraints. These constraints can be expressed as: In the formula, and Representing 1-norm and - Permissible deviation of probability under the norm condition. , The confidence level is expressed as the probability of uncertainty. The total number of scenes, To estimate the sample size for the photovoltaic prediction error distribution.

[0045] The constructed scheduling model is solved using the Column and Constraint Generation (CCG) algorithm. The optimal solution is gradually approximated through alternating iterations of the main problem (MP) and adversarial subproblems (SP). First, the MP is solved under the initial probability distribution to obtain start / stop and output decisions. Then, the SP is solved within the ambiguity set with these decisions fixed to obtain the worst-case probability distribution and corresponding cost. If the difference between the objective values ​​of the main and subproblems exceeds the tolerance, the worst-case distribution information is transformed into new constraints or columns and fed back to the MP, continuing the iteration until convergence. This method can effectively approximate the optimal robust scheduling solution under limited computational scale, balancing solution efficiency and robustness. Its solution process is mainly as follows: Main problem form: In the formula For the number of iterations, LP To estimate the lower bound of the optimal value of the original problem, This represents the probability of the scenario occurring.

[0046] Subproblem form: Let first , Then fix Solving the main problem yields Then update The value of makes At this point, the solution obtained is fixed. Solve the subproblem to obtain the new equation. Value, and update The value of makes To obtain new After finding the value, solve the main problem to obtain a new value. Value, and update , The value is repeated in this way until... .

[0047] First, the start / stop and output decisions are obtained by solving the maximum probability distribution (MP) under the initial probability distribution. Then, the minimum probability distribution (SP) is solved within the ambiguity set, fixing this decision, to obtain the worst-case probability distribution and its corresponding cost. If the difference between the objective values ​​of the main problem and the subproblems exceeds the tolerance, the worst-case distribution information is transformed into new constraints or columns and fed back to the MP, continuing the iteration until convergence. This method can effectively approximate the optimal robust scheduling solution under limited computational scale, balancing solution efficiency and robustness.

[0048] In summary, this method employs a long short-term memory neural network for photovoltaic (PV) output prediction, integrating historical data and real-time meteorological information to avoid the complexity of traditional physical modeling. A real-time error monitoring and adaptive model update mechanism is introduced to construct a PV digital twin model with real-time correction capabilities, improving prediction accuracy and long-term stability. Latin hypercube sampling and k-means clustering are combined to generate and reduce typical PV scenarios, enhancing the scheduling model's ability to represent uncertainty. Based on distributed robust optimization theory and a two-stage modeling method, the optimal scheduling strategy under the worst-case probability distribution is sought, balancing system economy and robustness. The column and constraint generation (CCG) algorithm is used for decomposition and solution, improving computational efficiency for large-scale integrated energy island scheduling problems in industrial parks while ensuring optimization accuracy.

[0049] Example 2 Based on Example 1, see Figure 4 This embodiment provides an energy management system for a park's integrated energy system that considers photovoltaic (PV) prediction errors. It implements the energy management method for the park's integrated energy system in Embodiment 1, integrating stochastic optimization and robust optimization principles to simultaneously consider the system's economy and robustness under uncertainty. Simultaneously, it uses Latin hypercube adoption and k-means clustering to generate initial PV output scenarios to describe the uncertainty of PV output. The optimal scheduling scheme for the park's integrated energy system is obtained by minimizing the sum of system operating costs and curtailment costs. The system includes: (1) Data acquisition module, used to acquire the actual output data of the photovoltaic array and the corresponding meteorological data, and to construct a training dataset.

[0050] (2) Data-driven model building module, used to build a data-driven prediction model based on neural network. The data-driven prediction model takes meteorological data as input and output data as output, and trains the data-driven prediction model based on the training dataset.

[0051] (3) Interaction and correction module, used to acquire real-time meteorological data and actual output data, use the trained data-driven prediction model to obtain predicted output data, calculate the output error between actual output data and predicted output data, determine whether the processing error meets the preset conditions, if not, adaptively update the data-driven prediction model, if so, according to the predicted output data and its output error distribution, extract multiple samples in the multidimensional random space through the Latin hypercube sampling method, generate the original scene set covering possible changes in photovoltaic output in various future periods, and reduce the scene through clustering.

[0052] (4) Decision module, which is used to construct a PIES distributed robust scheduling model under the constraints of the park's integrated energy system based on the reduced scenario set, solve the start-stop and output decisions, and realize the energy management of the park's integrated energy system.

[0053] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An energy management method for a park's integrated energy system that considers photovoltaic forecasting errors, characterized in that, Includes the following steps: Acquire the actual power output data of the photovoltaic array and the corresponding meteorological data to construct a training dataset; A data-driven prediction model based on a neural network is constructed. The data-driven prediction model takes meteorological data as input and output data as output, and is trained based on the training dataset. Acquire real-time meteorological data and actual power output data, use the trained data to drive the prediction model to obtain predicted power output data, and calculate the power output error between the actual power output data and the predicted power output data. Determine whether the processing error meets the preset conditions. If not, adaptively update the data-driven prediction model. If so, overlay the output error distribution on the predicted output data to generate an original scene set that covers possible changes in photovoltaic output in various future periods and conforms to the error distribution. Then, reduce the scene by clustering. Based on the reduced scenario set, a PIES distributed robust scheduling model is constructed under the constraints of the park's integrated energy system to solve start-up, shutdown, and power output decisions, thereby realizing energy management of the park's integrated energy system.

2. The energy management method for a comprehensive energy system in a park considering photovoltaic prediction errors according to claim 1, characterized in that, The PIES distributed robust scheduling model described above takes minimizing the system operating cost and the cost of light abandonment as its objective function.

3. The energy management method for a comprehensive energy system in a park considering photovoltaic prediction errors according to claim 1, characterized in that, The constraints of the park's integrated energy system include electrical power balance constraints, thermal power balance constraints, conventional unit constraints, and energy storage constraints. The conventional unit constraints include electrical power constraints, ramp-up constraints, and start-up / shutdown constraints. The energy storage constraints include upper and lower limit constraints and start-up / end-of-life constraints.

4. The energy management method for a comprehensive energy system in a park considering photovoltaic prediction errors according to claim 1, characterized in that, The PIES distributed robust scheduling model is as follows: in, , for , The feasible region consists of the corresponding constraints. These are deterministic quantities, unaffected by photovoltaic prediction errors, and include the start-up and shutdown status of conventional units and the charging and discharging status of electrical and thermal energy storage. This refers to the amount that is adjusted in real time according to changes in photovoltaic output, including the output value of each unit. For the cost of starting and stopping, For the cost of power generation and the cost of curtailment, The probability of the scenario occurring. Let be the set of feasible regions for the probability distribution of the scene. The total number of scenes, This is a transpose.

5. The energy management method for a comprehensive energy system in a park considering photovoltaic prediction errors according to claim 1, characterized in that, The solution process for the PIES distributed robust scheduling model includes the following steps: Under the constraints of 1-norm and ∞-norm, the start-stop and output decisions are obtained by solving the main problem based on the probability distribution of the initial output error. Under the premise of fixed start-stop and output decisions, the adversarial subproblem is solved in the ambiguity set to obtain the worst probability distribution and corresponding cost. If the difference between the objective values ​​of the main problem and the adversarial subproblem exceeds the tolerance, the worst distribution information is transformed into new constraints or columns and fed back to the main problem. The iteration continues until convergence.

6. The energy management method for a comprehensive energy system in a park considering photovoltaic prediction errors according to claim 5, characterized in that, The 1-norm and ∞-norm constraints are modeled as follows: In the formula, and Representing 1-norm and - Permissible deviation of probability under the norm condition. , The confidence level represents the probability of uncertainty. The total number of scenes, To estimate the sample size for the photovoltaic prediction error distribution, The main problem is modeled as follows: In the formula, , for , The feasible region consists of the corresponding constraints. These are deterministic quantities, unaffected by photovoltaic prediction errors, and include the start-up and shutdown status of conventional units and the charging and discharging status of electrical and thermal energy storage. This refers to the amount that is adjusted in real time according to changes in photovoltaic output, including the output value of each unit. For the cost of starting and stopping, For the cost of power generation and the cost of curtailment, The probability of the scenario occurring. The total number of scenes, For transpose, To estimate the lower bound of the optimal value of the original problem, This represents the current iteration number. The adversarial subproblem is modeled as follows: In the formula, Let be the set of feasible regions for the probability distribution of the scene. The optimal values ​​of the decision variables in the first stage, obtained from solving the main problem, are... For a given first-stage decision Below, the feasible region of the decision variables in the second stage.

7. The energy management method for a comprehensive energy system in a park considering photovoltaic prediction errors according to claim 1, characterized in that, The adaptive update includes incremental training, transfer learning, or parameter correction.

8. The energy management method for a comprehensive energy system in a park considering photovoltaic prediction errors according to claim 1, characterized in that, The data-driven prediction model is constructed based on long short-term memory neural networks, recurrent neural networks, or convolutional neural networks.

9. The energy management method for a comprehensive energy system in a park considering photovoltaic prediction errors according to claim 1, characterized in that, The meteorological data mentioned include solar irradiance, ambient temperature, and humidity.

10. An energy management system for a park integrated energy system that considers photovoltaic prediction errors, characterized in that, For implementing the energy management method of the integrated energy system of the park as described in any one of claims 1-9, the system includes: The data acquisition module is used to acquire the actual output data of the photovoltaic array and the corresponding meteorological data to build a training dataset; The data-driven model building module is used to build a data-driven prediction model based on a neural network. The data-driven prediction model takes meteorological data as input and output data as output, and is trained based on the training dataset. The interaction and correction module is used to acquire real-time meteorological data and actual output data, use the trained data to drive the prediction model to obtain predicted output data, calculate the output error between the actual output data and the predicted output data, and determine whether the processing error meets the preset conditions. If not, the data-driven prediction model is adaptively updated. If so, the output error distribution is superimposed on the predicted output data to generate an original scene set that covers possible changes in photovoltaic output in various future periods and conforms to the error distribution. Scene reduction is performed through clustering. The decision-making module is used to construct a PIES distributed robust scheduling model under the constraints of the park's integrated energy system based on the reduced scenario set, solve start-up and power output decisions, and realize energy management of the park's integrated energy system.

Images